Hierarchical Slow Feature Analysis and Top-Down Processes
-
1
Humboldt-Universität zu Berlin, Institute for Theoretical Biology, Germany
-
2
Humboldt-Universität zu Berlin, Bernstein Center for Computational Neuroscience, Germany
-
3
Ruhr-Universität Bochum, Institut für Neuroinformatik, Germany
Top-down processes are thought to play an important role in the mammalian visual system, e.g., for interpreting ambiguous stimuli. Slow Feature Analysis (SFA) [2] on the other hand is proven to be an efficient algorithm for the bottom-up processing of visual stimuli [2][3]. Therefore it seems natural to combine bottom-up SFA with top-down processes.
SFA is an unsupervised learning algorithm that leverages the time structure of incoming stimuli to extract higher-level features. The SFA algorithm works with continuous, real variables. The algorithm itself is linear, but can be combined with a prior expansion into a more powerful function space. Quadratic polynomials have been used successfully in hierarchical networks for the extraction of high-level features from complex visual stimuli. Unfortunately this expansion makes it difficult to relate input and output components in the layers. In particular it is generally not possible to invert the bottom-up mapping, which indicates serious obstacles for top-down processes. We explored techniques to address this inversion problem. Our methods combine gradient decent and vector quantization algorithms and allowed stimulus reconstruction at the lowest layer (see Fig. 1). The results also suggest that a further increase in reconstruction performance will require a different expansion that is partly optimized for the top-down step.
Figure 1. Stimulus reconstruction from higher-level features. (a) shows the reconstruction for a single receptive field patch on the lowest layer, with complex cell like output behavior. On the left is the original stimulus, on the right side the reconstruction, which was calculated from the layer output. In (b) the same reconstruction technique has been applied to a whole image. The first picture is the original stimulus, the second one is the reconstruction from the lowest layer output. The third image is the reconstruction from the second layer output, showing some significant reconstruction errors.
References
1. Wiskott L, Sejnowski TJ: Slow feature analysis: Unsupervised learning of invariances. Neural Computation 2002; 14(4):715-770.
2. Franzius M, Sprekeler H, and Wiskott L: Slowness and sparseness lead to place, head-diretion and spatial-view cells. Public Library of Science (PLoS) Computational Biology, 3(8):e166, 2007.
3. Franzius M, Wilbert N, and Wiskott L: Invariant Object Recognition with Slow Feature Analysis.Proc. 18th Int'l Conf. on Artificial Neural Networks, ICANN'08, Prague, September 3-6, eds. Vera Kurková and Roman Neruda and Jan Koutník, publ. Springer-Verlag, pp. 961-970.
Keywords:
computational neuroscience
Conference:
Bernstein Conference on Computational Neuroscience, Berlin, Germany, 27 Sep - 1 Oct, 2010.
Presentation Type:
Poster Abstract
Topic:
Bernstein Conference on Computational Neuroscience
Citation:
Wilbert
N and
Wiskott
L
(2010). Hierarchical Slow Feature Analysis and Top-Down Processes.
Front. Comput. Neurosci.
Conference Abstract:
Bernstein Conference on Computational Neuroscience.
doi: 10.3389/conf.fncom.2010.51.00119
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
09 Sep 2010;
Published Online:
23 Sep 2010.
*
Correspondence:
Dr. Niko Wilbert, Humboldt-Universität zu Berlin, Institute for Theoretical Biology, Berlin, Germany, mail@nikowilbert.de